Research On Updating Schemes For Erasure-coded In-memory Stores

Recently,erasure-coded in-memory stores have received more and more attention from researchers because of their low-response-delay and high-space-utilization advantages.Current researches on erasure-coded in-memory stores are mainly focused on read-only workloads,and they place emphasis on both in-memory data recovery and load balancing rather than data updating.In fact,the issue of data updating should be addressed for erasure-coded in-memory datasets.Generally,existing data updating schemes adopt the serial-updating manner,therefore the hardware performance of in-memory clusters can not be fully exploited.To address this problem,this paper studies data-updating schemes for erasure-coded in-memory clusters by incorporating parallel-updating mechanisms,thereby improving data-updating performance.In order to solve the low-updating-efficiency problem incurred by serial updating,a group-based parallel updating strategy called GU is proposed.The basic concept of GU strategy is to adopt an updating window for I/O requests and divide updating requests within an updating window into several groups.In particular,updating requests are separated into a certain number of groups in according to the location?i.e.,node?of data blocks to be updated,and all updating requests in a group can be concurrently executed.Different from serial updating which performs a single updating request at a time,GU-based parallel updating allows multiple updating requests in a group to be executed at the same time.On the basic of GU strategy,a hybrid updating scheme called Hybrid-U^[gu]is presented to handle a series of updating requests exhibiting various updating granularities.The core idea of Hybrid-U^[gu]scheme is to apply an appropriate updating method to updating requests in a group in accordance with the distribution of data blocks to be updated,thereby decreasing network traffic?i.e.,updating traffic?caused by updates.Specially,with GU strategy in place,such four updating methods as DUM,PUM,PUM-P,and PDN-P developed for erasure-coded storage clusters can be transformed into four corresponding updating methods for erasure-coded in-memory clusters,i.e.,DUM^[gu],PUM^[gu],PUM-P^[gu],and PDN-P^[gu].Thus,a Hybrid-U^[gu]scheme may be a combination of two or three GU-based updating methods,e.g.,‘DUM_[gu]+PUM-P_[gu]'.With replaying updating traces generated by the benchmark of YCSB upon the real-world in-memory store clusters,the above GU-based updating methods and Hybrid-U^[gu]schemes are quantitatively evaluated in terms of average updating time and updating traffic.Experimental results show that,compared to basic updating schemes,both GU-based updating methods and Hybrid-U^[gu]schemes can reduce average updating time by a factor of 53.6%;furthermore,compared to GU-based updating methods,the Hybrid-U^[gu]scheme can save updating traffic by 6%to 30%.